Prune bias from the root: Bias removal and fairness estimation by pruning sensitive attributes in pre-trained DNN models
Bibliographic record
Abstract
Deep learning models (DNNs) are widely used in high-stakes decision-making domains, but they often inherit and amplify biases present in training data, leading to unfair predictions. Given this context, fairness estimation metrics and bias removal methods are required to select and enhance fair models. However, we found that existing metrics lack robustness in estimating multi-attribute group fairness. Further, existing post-processing bias removal methods often focus on group fairness and fail to address individual fairness or optimize along multiple sensitive attributes. In this study, we explore the effectiveness of attribute pruning (i.e., zeroing out sensitive attribute weights in a pre-trained DNN’s input layer) in both bias removal and multi-attribute group fairness estimation. To study attribute pruning’s impact on bias removal, we conducted experiments on 32 models and 4 widely used datasets, and compared its effect in single-attribute group bias removal and accuracy preservation with 3 baseline post-processing methods. We then leveraged 3 datasets with multiple sensitive attributes to demonstrate how to use attribute pruning for multi-attribute group fairness estimation. Single-attribute pruning can better preserve model accuracy than conventional post-processing methods in 23 out of 32 cases, and enforces individual fairness by design. However, since individual fairness and group fairness are fundamentally different objectives, attribute pruning’s effect on group fairness metrics is often inconsistent. We also extend our approach to a multi-attribute setting, demonstrating its potential for improving individual fairness jointly across sensitive attributes and for enabling multi-attribute fairness-aware model selection. Attribute pruning is a practical post-processing approach for enforcing individual fairness, with limited and data-dependent impact on group fairness. These limitations reflect the inherent trade-off between individual and group fairness objectives. In addition, attribute pruning provides a useful mechanism for bias estimation, particularly in multi-attribute contexts. We advocate for its adoption as a comparison baseline in fairness-aware AI development and encourage further exploration. • Prove the effectiveness of single-attribute pruning for bias removal on DNNs. • Propose the use of attribute pruning on multiple features. • Introduce a robust multi-attribute fairness estimation method. • Explicitly discuss the limits of algorithmic bias removal and provide suggestions.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.013 | 0.038 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.002 | 0.001 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.002 | 0.001 |
| Scholarly communication | 0.002 | 0.003 |
| Open science | 0.002 | 0.003 |
| Research integrity | 0.003 | 0.007 |
| Insufficient payload (model declined to judge) | 0.002 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".